Cognitive bias in Facebook News Feed and Google Search
Facebook and Google are globally recognized companies, with billions in ad revenue each year. Recently, many people are worried they aren’t properly accounting for biases when ranking posts in Facebook News Feed or results in Google Search. This article gives an overview of the key events that transpired this year for News Feed and Search, closing with some steps they can start taking to reduce bias in ranking algorithms.
The evolution of News Feed
The News Feed algorithm has changed over time to reflect feedback from users. Since its conception, the focus has been on allowing users to find high quality, relevant content from pages while always prioritizing posts from friends and family.
However, users complained they were feeling disconnected from their News Feed since public content was muddling interactions from friends and family. So, in 2015, 2016, and 2017 Facebook continually made a variety of changes to their ranking algorithm to prioritize content you’d find relevant, based on who posted it, how often you interacted with them, and what topics were trending across all users. Despite these modifications, the focus remained on showing relevant content, where relevant was defined as something you’d engage with.
Then, at the beginning of this year, Mark Zuckerberg claims:
Recently we've gotten feedback from our community that public content -- posts from businesses, brands and media -- is crowding out the personal moments that lead us to connect more with each other.
Duh. Users had been complaining about this since 2013. He goes on further to explain a shift in strategy:
I'm changing the goal I give our product teams from focusing on helping you find relevant content to helping you have more wmeaningful social interactions.
I think this is a positive shift. The problem from the beginning was always this: what should Facebook optimize for? It's clear that users come to the News Feed to get updates from people they already know. On the other hand, businesses come to the News Feed to reach users they don’t know. In the middle is Facebook, struggling to balance the see-saw by prioritizing relevant content. Unfortunately, optimizing for relevant content, in this case, is equivalent to minimizing the user’s pain of seeing posts they don’t care about on their News Feed. By prioritizing meaningful social interactions, users will probably spend less time on the News Feed, but the time they spend will be more engaged and focused. I’m guessing Facebook believes this will increase ad conversion and avoid user churn.
Google Search in the spotlight
Sundar Pichai recently testified in front of the House Judiciary Committee. Several members of the Committee suggested that Google’s search results were biased against conservatives. Some went so far as to suggest the results were intentionally manipulated by Google or Google employees to support the company’s political agenda. A notable quote from Lamar Smith (R-TX):
Google has long-faced criticism for manipulating results to censor conservatives… PJ Media found that 96% of search results for Trump were from liberal media outlets. Not a single right-leaning site appeared on the first page of search results.
The methodology for the study that he’s referring to was as follows:
1. Type “Trump” into news.google.com.
2. Analyze the sources of results against a media bias chart.
Calling this a study is like calling a tweet a dissertation. Besides, the 96% statistic has been completely debunked by reputed researchers, and PJ Media’s supervising editor even agreed the study is “not scientific".
On the other hand, Darrell Issa (R-CA) made an interesting statement:
If you measure the outcome, what you find is there is an appearance of bias.
Put differently – if ranking algorithms are adapted based on user feedback, and users are biased, the algorithms become biased. Extreme cases of this cognitive bias in machine learning have occurred before, such as when Microsoft’s AI chatbot started spewing racist and inflammatory tweets it learned from other people on Twitter, or when Northpointe’s risk assessment score falsely labeled black defendants as future criminals twice as often as white defendants, or even when Google’s sentiment analyzer labeled statements such as “I am Jewish” as negative.
It’s likely that Google’s user base is already biased across a variety of dimensions – the digital marketing consultancy Further found in a 2015 research study that Google users are younger, more likely to be male, and more interested in gaming than Bing or Yahoo users. Google is heavily incentivized to optimize search results to be relevant to its current users since this results in higher engagement and more revenue. But doing so may create a feedback loop resulting in even more biased search results.
So what should they optimize for?
Facebook and Google have both optimized for showing relevant posts or search results. However, the News Feed still has to show ads, which are sometimes irrelevant. And Search users will find articles that validate their biases more relevant.
So what is the right thing to optimize for? Nobody really knows – this question has spawned a new field of research called machine bias.
One possible solution is to design recommendation algorithms to account for bias. This can be done proactively by using unbiased data to improve algorithms. It can also be done reactively by vetting results via third parties and modifying algorithms to disincentivize biased results. New research around “counterfactual fairness” attempts exactly this by negatively weighting algorithms that fail to account for biases in the data.
Another solution is to improve social awareness of how these ranking algorithms work. It’s clear from Zuckerberg and Pichai’s testimonies that we can’t yet rely on Capitol Hill to properly regulate these algorithms, seeing as they still don’t understand Facebook’s business model or Google’s search algorithm. While the onus should be on White House representatives to understand what they’re regulating, maybe Google and Facebook should take the time to explain and educate how these algorithms work. The current status quo of Capital Hill pointing fingers at Silicon Valley, pointing fingers back saying they’re ignorant won’t get us anywhere.
This is not a simple problem that we’re going to solve overnight. Accounting for bias in algorithms is extremely challenging – any algorithm that learns from human decisions will inherently be biased. However, there are steps we can take today to begin the journey towards reducing machine bias, and Google and Facebook, as thought leaders in the machine learning and consumer technology space, are the perfect role models to take this important first step.